[136] | 1 | // This file is part of Eigen, a lightweight C++ template library
|
---|
| 2 | // for linear algebra.
|
---|
| 3 | //
|
---|
| 4 | // Copyright (C) 2008 Konstantinos Margaritis <markos@codex.gr>
|
---|
| 5 | //
|
---|
| 6 | // This Source Code Form is subject to the terms of the Mozilla
|
---|
| 7 | // Public License v. 2.0. If a copy of the MPL was not distributed
|
---|
| 8 | // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
|
---|
| 9 |
|
---|
| 10 | #ifndef EIGEN_PACKET_MATH_ALTIVEC_H
|
---|
| 11 | #define EIGEN_PACKET_MATH_ALTIVEC_H
|
---|
| 12 |
|
---|
| 13 | namespace Eigen {
|
---|
| 14 |
|
---|
| 15 | namespace internal {
|
---|
| 16 |
|
---|
| 17 | #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
|
---|
| 18 | #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
|
---|
| 19 | #endif
|
---|
| 20 |
|
---|
| 21 | #ifndef EIGEN_HAS_FUSE_CJMADD
|
---|
| 22 | #define EIGEN_HAS_FUSE_CJMADD 1
|
---|
| 23 | #endif
|
---|
| 24 |
|
---|
| 25 | // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16
|
---|
| 26 | #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
|
---|
| 27 | #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16
|
---|
| 28 | #endif
|
---|
| 29 |
|
---|
| 30 | typedef __vector float Packet4f;
|
---|
| 31 | typedef __vector int Packet4i;
|
---|
| 32 | typedef __vector unsigned int Packet4ui;
|
---|
| 33 | typedef __vector __bool int Packet4bi;
|
---|
| 34 | typedef __vector short int Packet8i;
|
---|
| 35 | typedef __vector unsigned char Packet16uc;
|
---|
| 36 |
|
---|
| 37 | // We don't want to write the same code all the time, but we need to reuse the constants
|
---|
| 38 | // and it doesn't really work to declare them global, so we define macros instead
|
---|
| 39 |
|
---|
| 40 | #define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
|
---|
| 41 | Packet4f p4f_##NAME = (Packet4f) vec_splat_s32(X)
|
---|
| 42 |
|
---|
| 43 | #define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
|
---|
| 44 | Packet4i p4i_##NAME = vec_splat_s32(X)
|
---|
| 45 |
|
---|
| 46 | #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
|
---|
| 47 | Packet4f p4f_##NAME = pset1<Packet4f>(X)
|
---|
| 48 |
|
---|
| 49 | #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
|
---|
| 50 | Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
|
---|
| 51 |
|
---|
| 52 | #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
|
---|
| 53 | Packet4i p4i_##NAME = pset1<Packet4i>(X)
|
---|
| 54 |
|
---|
| 55 | #define DST_CHAN 1
|
---|
| 56 | #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
|
---|
| 57 |
|
---|
| 58 | // Define global static constants:
|
---|
| 59 | static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 };
|
---|
| 60 | static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 };
|
---|
| 61 | static Packet16uc p16uc_REVERSE = {12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3};
|
---|
| 62 | static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0);
|
---|
| 63 | static Packet16uc p16uc_DUPLICATE = {0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7};
|
---|
| 64 |
|
---|
| 65 | static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0);
|
---|
| 66 | static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0);
|
---|
| 67 | static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1);
|
---|
| 68 | static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16);
|
---|
| 69 | static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1);
|
---|
| 70 | static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0);
|
---|
| 71 | static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1);
|
---|
| 72 |
|
---|
| 73 | template<> struct packet_traits<float> : default_packet_traits
|
---|
| 74 | {
|
---|
| 75 | typedef Packet4f type;
|
---|
| 76 | enum {
|
---|
| 77 | Vectorizable = 1,
|
---|
| 78 | AlignedOnScalar = 1,
|
---|
| 79 | size=4,
|
---|
| 80 |
|
---|
| 81 | // FIXME check the Has*
|
---|
| 82 | HasSin = 0,
|
---|
| 83 | HasCos = 0,
|
---|
| 84 | HasLog = 0,
|
---|
| 85 | HasExp = 0,
|
---|
| 86 | HasSqrt = 0
|
---|
| 87 | };
|
---|
| 88 | };
|
---|
| 89 | template<> struct packet_traits<int> : default_packet_traits
|
---|
| 90 | {
|
---|
| 91 | typedef Packet4i type;
|
---|
| 92 | enum {
|
---|
| 93 | // FIXME check the Has*
|
---|
| 94 | Vectorizable = 1,
|
---|
| 95 | AlignedOnScalar = 1,
|
---|
| 96 | size=4
|
---|
| 97 | };
|
---|
| 98 | };
|
---|
| 99 |
|
---|
| 100 | template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; };
|
---|
| 101 | template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; };
|
---|
| 102 | /*
|
---|
| 103 | inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
|
---|
| 104 | {
|
---|
| 105 | union {
|
---|
| 106 | Packet4f v;
|
---|
| 107 | float n[4];
|
---|
| 108 | } vt;
|
---|
| 109 | vt.v = v;
|
---|
| 110 | s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
|
---|
| 111 | return s;
|
---|
| 112 | }
|
---|
| 113 |
|
---|
| 114 | inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
|
---|
| 115 | {
|
---|
| 116 | union {
|
---|
| 117 | Packet4i v;
|
---|
| 118 | int n[4];
|
---|
| 119 | } vt;
|
---|
| 120 | vt.v = v;
|
---|
| 121 | s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
|
---|
| 122 | return s;
|
---|
| 123 | }
|
---|
| 124 |
|
---|
| 125 | inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
|
---|
| 126 | {
|
---|
| 127 | union {
|
---|
| 128 | Packet4ui v;
|
---|
| 129 | unsigned int n[4];
|
---|
| 130 | } vt;
|
---|
| 131 | vt.v = v;
|
---|
| 132 | s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
|
---|
| 133 | return s;
|
---|
| 134 | }
|
---|
| 135 |
|
---|
| 136 | inline std::ostream & operator <<(std::ostream & s, const Packetbi & v)
|
---|
| 137 | {
|
---|
| 138 | union {
|
---|
| 139 | Packet4bi v;
|
---|
| 140 | unsigned int n[4];
|
---|
| 141 | } vt;
|
---|
| 142 | vt.v = v;
|
---|
| 143 | s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
|
---|
| 144 | return s;
|
---|
| 145 | }
|
---|
| 146 | */
|
---|
| 147 | template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) {
|
---|
| 148 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
|
---|
| 149 | float EIGEN_ALIGN16 af[4];
|
---|
| 150 | af[0] = from;
|
---|
| 151 | Packet4f vc = vec_ld(0, af);
|
---|
| 152 | vc = vec_splat(vc, 0);
|
---|
| 153 | return vc;
|
---|
| 154 | }
|
---|
| 155 |
|
---|
| 156 | template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) {
|
---|
| 157 | int EIGEN_ALIGN16 ai[4];
|
---|
| 158 | ai[0] = from;
|
---|
| 159 | Packet4i vc = vec_ld(0, ai);
|
---|
| 160 | vc = vec_splat(vc, 0);
|
---|
| 161 | return vc;
|
---|
| 162 | }
|
---|
| 163 |
|
---|
| 164 | template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); }
|
---|
| 165 | template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a) { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); }
|
---|
| 166 |
|
---|
| 167 | template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_add(a,b); }
|
---|
| 168 | template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_add(a,b); }
|
---|
| 169 |
|
---|
| 170 | template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_sub(a,b); }
|
---|
| 171 | template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_sub(a,b); }
|
---|
| 172 |
|
---|
| 173 | template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return psub<Packet4f>(p4f_ZERO, a); }
|
---|
| 174 | template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return psub<Packet4i>(p4i_ZERO, a); }
|
---|
| 175 |
|
---|
| 176 | template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
|
---|
| 177 | template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
|
---|
| 178 |
|
---|
| 179 | template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b,p4f_ZERO); }
|
---|
| 180 | /* Commented out: it's actually slower than processing it scalar
|
---|
| 181 | *
|
---|
| 182 | template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
|
---|
| 183 | {
|
---|
| 184 | // Detailed in: http://freevec.org/content/32bit_signed_integer_multiplication_altivec
|
---|
| 185 | //Set up constants, variables
|
---|
| 186 | Packet4i a1, b1, bswap, low_prod, high_prod, prod, prod_, v1sel;
|
---|
| 187 |
|
---|
| 188 | // Get the absolute values
|
---|
| 189 | a1 = vec_abs(a);
|
---|
| 190 | b1 = vec_abs(b);
|
---|
| 191 |
|
---|
| 192 | // Get the signs using xor
|
---|
| 193 | Packet4bi sgn = (Packet4bi) vec_cmplt(vec_xor(a, b), p4i_ZERO);
|
---|
| 194 |
|
---|
| 195 | // Do the multiplication for the asbolute values.
|
---|
| 196 | bswap = (Packet4i) vec_rl((Packet4ui) b1, (Packet4ui) p4i_MINUS16 );
|
---|
| 197 | low_prod = vec_mulo((Packet8i) a1, (Packet8i)b1);
|
---|
| 198 | high_prod = vec_msum((Packet8i) a1, (Packet8i) bswap, p4i_ZERO);
|
---|
| 199 | high_prod = (Packet4i) vec_sl((Packet4ui) high_prod, (Packet4ui) p4i_MINUS16);
|
---|
| 200 | prod = vec_add( low_prod, high_prod );
|
---|
| 201 |
|
---|
| 202 | // NOR the product and select only the negative elements according to the sign mask
|
---|
| 203 | prod_ = vec_nor(prod, prod);
|
---|
| 204 | prod_ = vec_sel(p4i_ZERO, prod_, sgn);
|
---|
| 205 |
|
---|
| 206 | // Add 1 to the result to get the negative numbers
|
---|
| 207 | v1sel = vec_sel(p4i_ZERO, p4i_ONE, sgn);
|
---|
| 208 | prod_ = vec_add(prod_, v1sel);
|
---|
| 209 |
|
---|
| 210 | // Merge the results back to the final vector.
|
---|
| 211 | prod = vec_sel(prod, prod_, sgn);
|
---|
| 212 |
|
---|
| 213 | return prod;
|
---|
| 214 | }
|
---|
| 215 | */
|
---|
| 216 | template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
|
---|
| 217 | {
|
---|
| 218 | Packet4f t, y_0, y_1, res;
|
---|
| 219 |
|
---|
| 220 | // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
|
---|
| 221 | y_0 = vec_re(b);
|
---|
| 222 |
|
---|
| 223 | // Do one Newton-Raphson iteration to get the needed accuracy
|
---|
| 224 | t = vec_nmsub(y_0, b, p4f_ONE);
|
---|
| 225 | y_1 = vec_madd(y_0, t, y_0);
|
---|
| 226 |
|
---|
| 227 | res = vec_madd(a, y_1, p4f_ZERO);
|
---|
| 228 | return res;
|
---|
| 229 | }
|
---|
| 230 |
|
---|
| 231 | template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
|
---|
| 232 | { eigen_assert(false && "packet integer division are not supported by AltiVec");
|
---|
| 233 | return pset1<Packet4i>(0);
|
---|
| 234 | }
|
---|
| 235 |
|
---|
| 236 | // for some weird raisons, it has to be overloaded for packet of integers
|
---|
| 237 | template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); }
|
---|
| 238 | template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
|
---|
| 239 |
|
---|
| 240 | template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
|
---|
| 241 | template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
|
---|
| 242 |
|
---|
| 243 | template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
|
---|
| 244 | template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
|
---|
| 245 |
|
---|
| 246 | // Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
|
---|
| 247 | template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
|
---|
| 248 | template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
|
---|
| 249 |
|
---|
| 250 | template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_or(a, b); }
|
---|
| 251 | template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
|
---|
| 252 |
|
---|
| 253 | template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); }
|
---|
| 254 | template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
|
---|
| 255 |
|
---|
| 256 | template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); }
|
---|
| 257 | template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); }
|
---|
| 258 |
|
---|
| 259 | template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
|
---|
| 260 | template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
|
---|
| 261 |
|
---|
| 262 | template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
|
---|
| 263 | {
|
---|
| 264 | EIGEN_DEBUG_ALIGNED_LOAD
|
---|
| 265 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
|
---|
| 266 | Packet16uc MSQ, LSQ;
|
---|
| 267 | Packet16uc mask;
|
---|
| 268 | MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword
|
---|
| 269 | LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword
|
---|
| 270 | mask = vec_lvsl(0, from); // create the permute mask
|
---|
| 271 | return (Packet4f) vec_perm(MSQ, LSQ, mask); // align the data
|
---|
| 272 |
|
---|
| 273 | }
|
---|
| 274 | template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
|
---|
| 275 | {
|
---|
| 276 | EIGEN_DEBUG_ALIGNED_LOAD
|
---|
| 277 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
|
---|
| 278 | Packet16uc MSQ, LSQ;
|
---|
| 279 | Packet16uc mask;
|
---|
| 280 | MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword
|
---|
| 281 | LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword
|
---|
| 282 | mask = vec_lvsl(0, from); // create the permute mask
|
---|
| 283 | return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data
|
---|
| 284 | }
|
---|
| 285 |
|
---|
| 286 | template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
|
---|
| 287 | {
|
---|
| 288 | Packet4f p;
|
---|
| 289 | if((ptrdiff_t(&from) % 16) == 0) p = pload<Packet4f>(from);
|
---|
| 290 | else p = ploadu<Packet4f>(from);
|
---|
| 291 | return vec_perm(p, p, p16uc_DUPLICATE);
|
---|
| 292 | }
|
---|
| 293 | template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from)
|
---|
| 294 | {
|
---|
| 295 | Packet4i p;
|
---|
| 296 | if((ptrdiff_t(&from) % 16) == 0) p = pload<Packet4i>(from);
|
---|
| 297 | else p = ploadu<Packet4i>(from);
|
---|
| 298 | return vec_perm(p, p, p16uc_DUPLICATE);
|
---|
| 299 | }
|
---|
| 300 |
|
---|
| 301 | template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
|
---|
| 302 | template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
|
---|
| 303 |
|
---|
| 304 | template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from)
|
---|
| 305 | {
|
---|
| 306 | EIGEN_DEBUG_UNALIGNED_STORE
|
---|
| 307 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
|
---|
| 308 | // Warning: not thread safe!
|
---|
| 309 | Packet16uc MSQ, LSQ, edges;
|
---|
| 310 | Packet16uc edgeAlign, align;
|
---|
| 311 |
|
---|
| 312 | MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword
|
---|
| 313 | LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword
|
---|
| 314 | edgeAlign = vec_lvsl(0, to); // permute map to extract edges
|
---|
| 315 | edges=vec_perm(LSQ,MSQ,edgeAlign); // extract the edges
|
---|
| 316 | align = vec_lvsr( 0, to ); // permute map to misalign data
|
---|
| 317 | MSQ = vec_perm(edges,(Packet16uc)from,align); // misalign the data (MSQ)
|
---|
| 318 | LSQ = vec_perm((Packet16uc)from,edges,align); // misalign the data (LSQ)
|
---|
| 319 | vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first
|
---|
| 320 | vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part
|
---|
| 321 | }
|
---|
| 322 | template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from)
|
---|
| 323 | {
|
---|
| 324 | EIGEN_DEBUG_UNALIGNED_STORE
|
---|
| 325 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
|
---|
| 326 | // Warning: not thread safe!
|
---|
| 327 | Packet16uc MSQ, LSQ, edges;
|
---|
| 328 | Packet16uc edgeAlign, align;
|
---|
| 329 |
|
---|
| 330 | MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword
|
---|
| 331 | LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword
|
---|
| 332 | edgeAlign = vec_lvsl(0, to); // permute map to extract edges
|
---|
| 333 | edges=vec_perm(LSQ, MSQ, edgeAlign); // extract the edges
|
---|
| 334 | align = vec_lvsr( 0, to ); // permute map to misalign data
|
---|
| 335 | MSQ = vec_perm(edges, (Packet16uc) from, align); // misalign the data (MSQ)
|
---|
| 336 | LSQ = vec_perm((Packet16uc) from, edges, align); // misalign the data (LSQ)
|
---|
| 337 | vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first
|
---|
| 338 | vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part
|
---|
| 339 | }
|
---|
| 340 |
|
---|
| 341 | template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
|
---|
| 342 | template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
|
---|
| 343 |
|
---|
| 344 | template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
|
---|
| 345 | template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
|
---|
| 346 |
|
---|
| 347 | template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); }
|
---|
| 348 | template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); }
|
---|
| 349 |
|
---|
| 350 | template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); }
|
---|
| 351 | template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
|
---|
| 352 |
|
---|
| 353 | template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
|
---|
| 354 | {
|
---|
| 355 | Packet4f b, sum;
|
---|
| 356 | b = (Packet4f) vec_sld(a, a, 8);
|
---|
| 357 | sum = vec_add(a, b);
|
---|
| 358 | b = (Packet4f) vec_sld(sum, sum, 4);
|
---|
| 359 | sum = vec_add(sum, b);
|
---|
| 360 | return pfirst(sum);
|
---|
| 361 | }
|
---|
| 362 |
|
---|
| 363 | template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
|
---|
| 364 | {
|
---|
| 365 | Packet4f v[4], sum[4];
|
---|
| 366 |
|
---|
| 367 | // It's easier and faster to transpose then add as columns
|
---|
| 368 | // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
|
---|
| 369 | // Do the transpose, first set of moves
|
---|
| 370 | v[0] = vec_mergeh(vecs[0], vecs[2]);
|
---|
| 371 | v[1] = vec_mergel(vecs[0], vecs[2]);
|
---|
| 372 | v[2] = vec_mergeh(vecs[1], vecs[3]);
|
---|
| 373 | v[3] = vec_mergel(vecs[1], vecs[3]);
|
---|
| 374 | // Get the resulting vectors
|
---|
| 375 | sum[0] = vec_mergeh(v[0], v[2]);
|
---|
| 376 | sum[1] = vec_mergel(v[0], v[2]);
|
---|
| 377 | sum[2] = vec_mergeh(v[1], v[3]);
|
---|
| 378 | sum[3] = vec_mergel(v[1], v[3]);
|
---|
| 379 |
|
---|
| 380 | // Now do the summation:
|
---|
| 381 | // Lines 0+1
|
---|
| 382 | sum[0] = vec_add(sum[0], sum[1]);
|
---|
| 383 | // Lines 2+3
|
---|
| 384 | sum[1] = vec_add(sum[2], sum[3]);
|
---|
| 385 | // Add the results
|
---|
| 386 | sum[0] = vec_add(sum[0], sum[1]);
|
---|
| 387 |
|
---|
| 388 | return sum[0];
|
---|
| 389 | }
|
---|
| 390 |
|
---|
| 391 | template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
|
---|
| 392 | {
|
---|
| 393 | Packet4i sum;
|
---|
| 394 | sum = vec_sums(a, p4i_ZERO);
|
---|
| 395 | sum = vec_sld(sum, p4i_ZERO, 12);
|
---|
| 396 | return pfirst(sum);
|
---|
| 397 | }
|
---|
| 398 |
|
---|
| 399 | template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
|
---|
| 400 | {
|
---|
| 401 | Packet4i v[4], sum[4];
|
---|
| 402 |
|
---|
| 403 | // It's easier and faster to transpose then add as columns
|
---|
| 404 | // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
|
---|
| 405 | // Do the transpose, first set of moves
|
---|
| 406 | v[0] = vec_mergeh(vecs[0], vecs[2]);
|
---|
| 407 | v[1] = vec_mergel(vecs[0], vecs[2]);
|
---|
| 408 | v[2] = vec_mergeh(vecs[1], vecs[3]);
|
---|
| 409 | v[3] = vec_mergel(vecs[1], vecs[3]);
|
---|
| 410 | // Get the resulting vectors
|
---|
| 411 | sum[0] = vec_mergeh(v[0], v[2]);
|
---|
| 412 | sum[1] = vec_mergel(v[0], v[2]);
|
---|
| 413 | sum[2] = vec_mergeh(v[1], v[3]);
|
---|
| 414 | sum[3] = vec_mergel(v[1], v[3]);
|
---|
| 415 |
|
---|
| 416 | // Now do the summation:
|
---|
| 417 | // Lines 0+1
|
---|
| 418 | sum[0] = vec_add(sum[0], sum[1]);
|
---|
| 419 | // Lines 2+3
|
---|
| 420 | sum[1] = vec_add(sum[2], sum[3]);
|
---|
| 421 | // Add the results
|
---|
| 422 | sum[0] = vec_add(sum[0], sum[1]);
|
---|
| 423 |
|
---|
| 424 | return sum[0];
|
---|
| 425 | }
|
---|
| 426 |
|
---|
| 427 | // Other reduction functions:
|
---|
| 428 | // mul
|
---|
| 429 | template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
|
---|
| 430 | {
|
---|
| 431 | Packet4f prod;
|
---|
| 432 | prod = pmul(a, (Packet4f)vec_sld(a, a, 8));
|
---|
| 433 | return pfirst(pmul(prod, (Packet4f)vec_sld(prod, prod, 4)));
|
---|
| 434 | }
|
---|
| 435 |
|
---|
| 436 | template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
|
---|
| 437 | {
|
---|
| 438 | EIGEN_ALIGN16 int aux[4];
|
---|
| 439 | pstore(aux, a);
|
---|
| 440 | return aux[0] * aux[1] * aux[2] * aux[3];
|
---|
| 441 | }
|
---|
| 442 |
|
---|
| 443 | // min
|
---|
| 444 | template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
|
---|
| 445 | {
|
---|
| 446 | Packet4f b, res;
|
---|
| 447 | b = vec_min(a, vec_sld(a, a, 8));
|
---|
| 448 | res = vec_min(b, vec_sld(b, b, 4));
|
---|
| 449 | return pfirst(res);
|
---|
| 450 | }
|
---|
| 451 |
|
---|
| 452 | template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
|
---|
| 453 | {
|
---|
| 454 | Packet4i b, res;
|
---|
| 455 | b = vec_min(a, vec_sld(a, a, 8));
|
---|
| 456 | res = vec_min(b, vec_sld(b, b, 4));
|
---|
| 457 | return pfirst(res);
|
---|
| 458 | }
|
---|
| 459 |
|
---|
| 460 | // max
|
---|
| 461 | template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
|
---|
| 462 | {
|
---|
| 463 | Packet4f b, res;
|
---|
| 464 | b = vec_max(a, vec_sld(a, a, 8));
|
---|
| 465 | res = vec_max(b, vec_sld(b, b, 4));
|
---|
| 466 | return pfirst(res);
|
---|
| 467 | }
|
---|
| 468 |
|
---|
| 469 | template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
|
---|
| 470 | {
|
---|
| 471 | Packet4i b, res;
|
---|
| 472 | b = vec_max(a, vec_sld(a, a, 8));
|
---|
| 473 | res = vec_max(b, vec_sld(b, b, 4));
|
---|
| 474 | return pfirst(res);
|
---|
| 475 | }
|
---|
| 476 |
|
---|
| 477 | template<int Offset>
|
---|
| 478 | struct palign_impl<Offset,Packet4f>
|
---|
| 479 | {
|
---|
| 480 | static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
|
---|
| 481 | {
|
---|
| 482 | if (Offset!=0)
|
---|
| 483 | first = vec_sld(first, second, Offset*4);
|
---|
| 484 | }
|
---|
| 485 | };
|
---|
| 486 |
|
---|
| 487 | template<int Offset>
|
---|
| 488 | struct palign_impl<Offset,Packet4i>
|
---|
| 489 | {
|
---|
| 490 | static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
|
---|
| 491 | {
|
---|
| 492 | if (Offset!=0)
|
---|
| 493 | first = vec_sld(first, second, Offset*4);
|
---|
| 494 | }
|
---|
| 495 | };
|
---|
| 496 |
|
---|
| 497 | } // end namespace internal
|
---|
| 498 |
|
---|
| 499 | } // end namespace Eigen
|
---|
| 500 |
|
---|
| 501 | #endif // EIGEN_PACKET_MATH_ALTIVEC_H
|
---|